Image quality improvement with deep learning‐based reconstruction on abdominal ultrahigh‐resolution CT: A phantom study

Abstract Purpose In an ultrahigh‐resolution CT (U‐HRCT), deep learning‐based reconstruction (DLR) is expected to drastically reduce image noise without degrading spatial resolution. We assessed a new algorithm's effect on image quality at different radiation doses assuming an abdominal CT protocol. Methods For the normal‐sized abdominal models, a Catphan 600 was scanned by U‐HRCT with 100%, 50%, and 25% radiation doses. In all acquisitions, DLR was compared to model‐based iterative reconstruction (MBIR), filtered back projection (FBP), and hybrid iterative reconstruction (HIR). For the quantitative assessment, we compared image noise, which was defined as the standard deviation of the CT number, and spatial resolution among all reconstruction algorithms. Results Deep learning‐based reconstruction yielded lower image noise than FBP and HIR at each radiation dose. DLR yielded higher image noise than MBIR at the 100% and 50% radiation doses (100%, 50%, DLR: 15.4, 16.9 vs MBIR: 10.2, 15.6 Hounsfield units: HU). However, at the 25% radiation dose, the image noise in DLR was lower than that in MBIR (16.7 vs. 26.6 HU). The spatial frequency at 10% of the modulation transfer function (MTF) in DLR was 1.0 cycles/mm, slightly lower than that in MBIR (1.05 cycles/mm) at the 100% radiation dose. Even when the radiation dose decreased, the spatial frequency at 10% of the MTF of DLR did not change significantly (50% and 25% doses, 0.98 and 0.99 cycles/mm, respectively). Conclusion Deep learning‐based reconstruction performs more consistently at decreasing dose in abdominal ultrahigh‐resolution CT compared to all other commercially available reconstruction algorithms evaluated.


| INTRODUCTION
An ultrahigh-resolution CT (U-HRCT) scanner became available for clinical practice in 2017, and several studies have reported its advantages. [1][2][3][4][5][6] However, the increased image noise that occurs when U-HRCT is applied (along with the improved spatial resolution) is a common concern, 1,3-8 because the amount of image noise is affected by the slice thickness and matrix size. Abdominal dynamic CT is an essential imaging modality for malignant liver tumors such as hepatocellular carcinoma, cholangiocellular carcinoma, and metastatic tumors. 9-11 U-HRCT is expected to demonstrate tiny vessels and pathological conditions in greater detail. With U-HRCT, it is difficult to successfully control the appropriate radiation dose because the tube current is limited by the combination of focal spot size and exposure time. In addition, for multiphase scanning, it is necessary to minimize the radiation dose in each phase while maintaining a diagnostically adequate image quality. [12][13][14][15] Although model-based iterative reconstruction (MBIR) is used to reduce the image noise that accompanies an insufficient radiation dose, the use of MBIR requires considerable computational time for image reconstruction, 16 which can affect clinical practice.
A deep learning-based reconstruction (DLR) algorithm was recently released in U-HRCT and is expected to reduce image noise dramatically without degrading spatial resolution. 16,17 Compared to MBIR, DLR can reconstruct images more quickly and is expected to minimize a change in noise texture that is specific to the iterative reconstruction derived from a low radiation dose or the level of iterative reconstruction. [18][19][20] Therefore, we evaluated DLR in abdominal dynamic CT using U-HRCT. In the present study, we assessed the image noise and spatial resolution characteristics of the DLR algorithm at different radiation doses on abdominal U-HRCT compared with filtered back projection (FBP), hybrid iterative reconstruction (HIR), and MBIR.

| ME TH ODS
Our present study was performed with a phantom imaging experiment. Therefore, there was no need for institutional review board approval.

2.A | DLR algorithm
DLR incorporates a deep convolutional neural networks (DCNN) restoration process into the reconstruction flow. For the deep learning-based approach, given HIR images and high-dose MBIR images as training pairs, statistical features that differentiate signal from the noise and artifacts could be "learned" in the training process and then be "updated" in the DCNN kernel for future inference use. Millions of image pairs were used in the training of DLR. The gold standard clinical reference images were acquired with high tube current and reconstructed with true MBIR, and the true MBIR used a greater number of iterations than could be otherwise used in a clinical setting due to time constraints. No phantom data were included in the gold standard reference images. This training process was previously completed during the development phase with no off-site unsupervised training-which could alter the algorithm performance-taking place.

2.C | CT scanning
The phantom was scanned using the U-HRCT scanner (Aquilion Precision; Canon Medical Systems, Otawara, Japan). This scanner has three scan modes: normal-, high-, and superhigh-resolution modes. 5 In the present study, we used the high-resolution mode with an 80row detector configuration of 0.5 mm detectors (1792 channels).
Image reconstruction was performed with a 1024 × 1024 matrix size. The nominal focal spot size of the X-ray tube was 0.9 × 1.2 mm, which was equivalent to the small focus of the conventional area detector CT (Aquilion ONE ViSION edition; Canon Medical Systems, Otawara, Japan). The combination of scan mode and focal spot size was selected assuming abdominal dynamic CT because the selected focal spot size limits the maximum tube current and exposure time. The other imaging parameters were as follows: Tube voltage was set to 120 kVp and the rotation time was set to 0.5 s. Tube current was varied from 590 to 300 to 150 mA as F I G . 1. Axial image of the body phantom. A Catphan phantom attached to an oval annulus (25 × 35 cm; 95-cm circumference). The CTP 404, 486, and 515 modules were used to assess the in-plane spatial resolution, the image noise, and the low contrast detectability, respectively. The high-(estimated value of 340 HU), moderate-(estimated value of 120 HU), and low-(estimated value of −35 HU) signal objects were located at 9, 11, and 1 o'clock, respectively. SHIRASAKA ET AL. | 287 100%, 50%, and 25% radiation doses, respectively. Images were reconstructed at a 0.5-mm thickness with a 400 mm of field of view.
For evaluation of low contrast detectability, images were reconstructed at a 5-mm thickness. In all acquisitions, DLR (Advanced Intelligent Clear-IQ Engine [AiCE], Canon Medical Systems) with a clinically optimized body parameter, "body standard" and FBP reconstruction with the FC13 kernel were performed. In addition, two types of iterative reconstruction were conducted: HIR with the FC13 kernel, and MBIR reconstruction (AIDR 3D standard and FIRST body standard, respectively; Canon Medical Systems) ( Table 1).

2.D | Image assessment
For the quantitative assessment, the CT number, image noise, frequency characteristics of the image noise, signal visibility, and spatial resolution in all of the reconstruction algorithms were compared using Image J 1.52a (National Institutes of Health, Bethesda, MD, USA) and Excel 2016 (Microsoft, Redmond, WA). These quantitative analyses were obtained from one scan series. For the assessment of the low contrast detectability, a visual evaluation was performed by two radiologists.

2.D.1 | CT number
The CT numbers obtained with each reconstruction algorithm in the phantom experiments were compared using the CTP 404 module. 21 Three different disk-shaped objects with a diameter of 12 mm were used as the assumed abdominal structures (lumen of a contrastenhanced aorta, contrast-enhanced tumor, and adipose tissue 22 ). We used a delrin rod (estimated value of 340 Hounsfield units: HU), an acrylic rod (estimated value of 120 HU), and a polystyrene rod (estimated value of −35 HU) as the high-, moderate-, and low-signal objects, respectively (Fig. 1). The CT numbers of these objects for each reconstruction algorithm were recorded radially around the object centers at 1°intervals using a line region of interest (ROI) (1 × 27 pixels, 5 mm circle-radius). Mean CT numbers were obtained using the individual line ROI measurements [ Fig. 2(a)].

2.D.2 | Image noise magnitude
The standard deviation (SD) of the CT number was defined as the image noise in the axial image. A square ROI (256 × 256 pixels) was placed at the center of each axial image of the CTP 486 module. We calculated the mean image noise magnitude using 50 sequential images.

2.D.3 | Noise power spectrum
To evaluate the frequency characteristics of the image noise, we calculated the noise power spectrum (NPS) by the radial frequency method 23 using the CTP 486 module. The NPS curve was obtained from the center (256 × 256 pixels) ROIs used to analyze image noise. We also normalized the NPS by dividing the NPS value by the area-under-the-curve of the NPS.

2.D.4 | Low contrast detectability
Two board-certified radiologists (Y.U. and A.N.) with 22 and 27 years of experience in abdominal radiology, respectively, and blinded to the radiation dose and reconstruction method independently evaluated the low contrast detectability using axial images of the CTP 515 module (Fig. 3). The diameters of the low contrast object at the 1.0% contrast level were 15,9,8,7,6,5,4,3, and 2 mm. Each observer recorded the detectable minimum diameter of the low contrast object. Sixty images (three radiation doses × five scans × four types of reconstruction algorithms) were presented in random order to the two observers. Images were displayed with a window level and width of 60 and 250 HU, respectively. The low contrast detectability for each reconstruction algorithm was the median of the diameters reported by the two observers.

2.D.5 | Signal visibility
For the evaluation of the visibility of signals on axial images, attenuation profile curves (APCs) for all reconstruction algorithms were compared using the CTP 404 module. To obtain the mean APCs of the high-, moderate-, and low-signal objects for each reconstruction algorithm, we recorded 360 APCs radially around the object's center at 1°intervals using a line ROI [1 × 51 pixels, 10 mm circle-radius; Fig. 2  were calculated using an inserted disk-shaped object (Teflon, estimated value of 990 HU) surrounded by a square ROI according to the disk methodology. 24 First, a signal-averaging image was generated from the 50 sequential images to reduce image noise. Ten consecutive signal-averaging images reconstructed at 0.1-mm intervals were used to obtain the mean MTF value. Then, the edge of the object was analyzed to determine the edge-spread function, which was differentiated to obtain the line-spread function. Finally, an object-specific MTF was generated by Fourier transformation of the line-spread function.      3.B | Noise power spectrum F I G . 4. The clipped axial images (CTP404 module) of each reconstruction algorithm obtained at the 100%, 50%, and 25% radiation doses. As assumed abdominal structures (lumen of a contrast-enhanced aorta, contrast-enhanced tumor, and adipose tissue), three different disk-shaped objects (high, moderate, and low) with a 12-mm dia. were used to evaluate the signal visibility.

3.A | CT number and image noise magnitude
T A B L E 2 CT number in each reconstruction algorithm at different radiation doses.  Table 4 shows the detectable minimum diameter of each reconstruction algorithm at each relative radiation dose. As the radiation dose decreased, the detectable minimum diameters of the low-contrast object increased at all reconstruction algorithms. A substantial interrater agreement was observed (k = 0.713). Figure 7 shows the clipped axial images of each reconstruction algorithm obtained at the different radiation doses.

3.D | Attenuation profile curves
For the high-signal objects, the APC of each reconstruction algorithm except for FBP was somewhat consistent within the signal test object under the lower radiation doses [ Fig. 8 3.E | Spatial resolution Table 5 shows the spatial frequency at 50% and 10% of MTF in each reconstruction algorithm generated with the different radiation doses.
The spatial frequencies at 50% and 10% of the MTF of DLR scanned with the 100% radiation dose were 0.62 and 1.00 cycles/mm, respectively. When the radiation dose was decreased, the spatial frequency at the 50% value of the MTF of DLR declined slightly, but that of the  10% value did not change significantly (50% and 25% doses, 0.98 and 0.99 cycles/mm, respectively).

| DISCUSSION
The results of these experiments demonstrated that the DLR maintained the same image noise magnitude and spatial resolution at all radiation doses with less change in signal visibility than the commercially available reconstruction methods analyzed. The DLR algorithm may provide image quality benefits to U-HRCT platforms over those of the other techniques. It is notable that the image noise in the DLR was relatively consistent across a wide range of radiation doses. In the comparison of the DLR and MBIR algorithms, the image noise was higher in DLR than in MBIR at the highest radiation dose. However, at the lowest radiation dose, the image noise in DLR was lower than that in MBIR. We speculate that the MBIR algorithm might sacrifice de-noising in order to maintain the spatial resolution at lower radiation doses, whereas DLR provided a better tradeoff in terms of noise versus dose. Therefore, for the range of radiation doses investigated, the DLR algorithm can improve image noise performance at low CT doses, which is critical for the larger matrix utilized in U-HRCT.
The frequency characteristics of the image noise in DLR were similar to those in HIR although the training image of DLR is generated using a high-dose MBIR image. The HIR image is reconstructed first and used internally as an input image for the DLR image process. 16 Therefore, the shape of the NPS of the DLR image might be similar to those of HIR.
The shape of the normalized NPS for DLR minimally varied at the different radiation doses, while that of MBIR shifted slightly to a higher frequency as the radiation dose decreased. It is notable that the normalized NPS of DLR did not vary according to the radiation dose, and thus DLR can provide de-noising without a change in noise texture compared to MBIR. The image noise frequency characteristics in MBIR observed in this study contrasted with findings of previous studies. 18,19 We believe these differences may be due to photon starvation from the low doses utilized in this study, the number of channels in the CT system (conventional CT, U-HRCT: approximately 890 channels, 1792 channels, respectively), and the increased phantom size (circular phantoms of 15.0 and 21.5 cm diameter, in previous studies, versus an oval phantom of 25 × 35 cm in the present study). The MBIR performance in this study appears be much more dose-dependent than DLR performance. Therefore, DLR appears more likely to provide superior reconstruction capabilities under the necessary low-dose conditions of U-HRCT. We thus consider DLR to be useful for dynamic abdominal CT on U-HRCT, which tends to have excessive noise.

| CONCLUSIONS
The present quantitative evaluations showed that the DLR performs more consistently at decreasing dose than MBIR, HIR, or FBP without extraordinary compromises in spatial resolution and low contrast detectability as compared with other reconstruction algorithms, and without a significant computation penalty. In particular, at lower radiation doses, DRL quantitatively performed better than MBIR and is expected to reduce image noise. For abdominal dynamic CT on U-HRCT, DLR may be a promising tool to compensate for the increased image noise from smaller detectors in a larger matrix.

D A T A A V A I L A B I L I T Y S T A T E M E N T
Research data are not shared.

R E F E R E N C E S
T A B L E 5 Spatial frequency at 50% and 10% of MTF at different radiation doses.